https://x.com/jandotai/status/1915670642139226319?t=kUirBijj0CTMNkjxlmy2Bw&s=35
X (formerly Twitter)
👋 Jan (@jandotai) on X
Tina: 1.5B model that beats other 1.5B reasoning models trained with full-parameter RL, at just $9 total cost.
It hits 43.33% on AIME24, outperforming models trained with 20-40x more compute.
Post-trained from R1's Qwen-1.5B distill using LoRA and RL. …
It hits 43.33% on AIME24, outperforming models trained with 20-40x more compute.
Post-trained from R1's Qwen-1.5B distill using LoRA and RL. …
